Non-stationarity that breaks the Markov assumption followed by most single-agent reinforcement learning algorithms is one of the main challenges in multi-agent environment， making each agent may be caught in an infinite loop caused by the environment created by the other agents during the learning process. To solve above problem， the implementation method of Centralized Training with Decentralized Execution （CTDE） structure in reinforcement learning was studied， and from two perspectives of agent communication and exploration， the QMIX algorithm was improved by introducing a Variance Control-Based （VBC） communication model and a curiosity mechanism. The proposed algorithm was validated in micro control scenarios of StarCraft Ⅱ Learning Environment （SC2LE）. Experimental results show that the proposed algorithm can improve the performance and obtain a training model with higher convergence speed compared to QMIX algorithm.

Software defect prediction is a hot research topic in the field of software quality assurance. The quality of defect prediction models is closely related to the training data. The datasets used for defect prediction mainly have the problems of data feature selection and data class imbalance. Aiming at the problem of data feature selection, common process features of software development and the newly proposed extended process features were used, and then the feature selection algorithm based on clustering analysis was used to perform feature selection. Aiming at the data class imbalance problem, an improved Borderline-SMOTE (Borderline-Synthetic Minority Oversampling Technique) method was proposed to make the numbers of positive and negative samples in the training dataset relatively balanced, and make the characteristics of the synthesized samples more consistent with the actual sample characteristics. Experiments were performed by using the open source datasets of projects such as bugzilla and jUnit. The results show that the used feature selection algorithm can reduce the model training time by 57.94% while keeping high F-measure value of the model; compared to the defect prediction model obtained by using the original method to process samples, the model obtained by the improved Borderline-SMOTE method respectively increase the Precision, Recall, F-measure, and AUC (Area Under the Curve) by 2.36 percentage points, 1.8 percentage points, 2.13 percentage points and 2.36 percentage points on average; the defect prediction model obtained by introducing the extended process features has an average improvement of 3.79% in F-measure value compared to the model without the extended process features; compared with the models obtained by methods in the literatures, the model obtained by the proposed method has an average increase of 15.79% in F-measure value. The experimental results prove that the proposed method can effectively improve the quality of the defect prediction model.

To suppress judgment one-sidedness in the existing bearing fault diagnosis method, a bearing fault diagnosis method based on Gibbs sampling was proposed. Firstly, the bearing vibration signal was decomposed by Local Characteristic Scale Decomposition (LCD) to obtain Intrinsic Scale Components (ISC). Secondly, the time domain features were extracted from the bearing vibration signal and ISC, and the time domain features were ranked according to feature sensitivity level. The top ranked features were selected to make up feature sets. Thirdly, feature set training was used to generate a multi-dimensional Gaussian distribution model based on Gibbs sampling. Finally, posterior analysis was used to obtain the probability to realize bearing fault diagnosis. The experimental results show that the diagnostic accuracy of the proposed method reaches 100%; compared with the bearing diagnosis method based on SVM (Support Vector Machine), the diagnostic accuracy is improved by 11.1 percentage points when the number of features is 43. The proposed method can effectively diagnose rolling bearing faults, and it also has good diagnostic effect on high-dimensional and complex bearing fault data.

When the appearance of an object changes rapidly, most of the weak learners can not capture the new feature distributions which will lead to tracking failure. In order to deal with that issue, a Gaussian weighted online multiple classifiers algorithm boosting for object tracking was proposed. This algorithm defined one weak classifier which included a simple visual feature and a threshold for each domain problem. Gaussian weighting function was introduced to weigh each weak classifier's contribution in a particular sample, therefore the tracking performance was improved through joint learning of multiple classifiers. In the process of object tracking, online multiple classifiers can not only simultaneously determine the location and estimate the pose of the object, but also successfully learn multi-modal appearance models and track an object under rapid appearance changes. The experimental results show that, after a short initial training phase, the average tracking error rate of the proposed algorithm is 12.8%, which proves that the tracking performance has enhanced significantly.

Since the traditional kernel fuzzy clustering algorithm does not take account of the level of contribution the every feature makes toward clustering, and it also has a shortcoming of easy falling into a situation of a local optimum, an improved kernel fuzzy clustering algorithm was put forward. It combined the advantage of the global optimum to construct a simple and effective fitness function that can avoid plunging local optimum. This improved algorithm gave every feature a weighted coefficient, in which the ReliefF algorithm was used to assign the weights for every feature. Compared with the traditional algorithm, this one has made some significant progress, and the experimental result has proved its effectiveness.